Академический Документы
Профессиональный Документы
Культура Документы
1
SUPRAS & GRASP, B5, University of Liège, B-4000 Liège, Belgium
2
Pennsylvania State University, University Park PA 16802, USA
were already closed [2]. A mapping of the Nikkei, DAX and DJIA daily sign
fluctuations has been made onto a 1/2 Ising spin chain as if there was a
continuous index calculated three times during 24 hours. This showed that
the spin cluster fluctuations are rather equivalent to random fluctuations, -
except during pre-crash periods in which (down − down − down) spin clus-
ters form with a higher probability than expected if the fluctuations are to
be considered independent of each other [3]. This has allowed to eliminate
all criticisms about the major responsibility of the derivative markets in the
United States on the Oct. 87 crash. Since then, the world-wide interdepen-
dence of the economy has been going on still more strongly. Thus if really
a speculative bubble is occurring on some financial markets, as commonly
observed in the recent years, the phenomenon of propagation to other stock
exchanges could be more important now. It is known that methods of negoti-
ation have widely changed on many financial places. Markets using electronic
systems of negotiation take advantage of recent improvement in their trans-
action capacity. It is easier today to face a substantial increase in transaction
volumes during a major crisis period. Moreover, the efficient use of the deriva-
tive markets could avoid useless pressures on the traditional market. These
financial factual observations should be turned into quantitative measures,
in order to, if necessary, avoid crashes.
Whence there is a need for techniques capable of rapidly following a bub-
ble explosion or preventing it. Notice that the drop of stock market indices
can not only spread out over two or three days but also over a much longer
period. The example of the Tokyo stock exchange at the beginning of the
1990’s is a prominent illustration. By comparison to the most famous crash
of 1929, the Oct. 87 crash was spread over 2 days: the Dow Jones sank 12.8 %
on October 28 and 11.7 % on the following day. (That was similar for the
DAX which dropped by 8.00 % and 7.00 % on Oct. 26 and 28, 1987 respec-
tively.) This shows that a stock market index decline does not necessarily
lead to a crash in one single day. Indeed, the decline can be slow and last
several days or even several months in what would be called not a crash, but
a long duration bear market.
In the present econophysics research context, it is of interest to examine
whether the evolution of quotations on the main stock exchange places have
similarities and whether crash symptoms can be found. Even if history gen-
erally tends to repeat, does it always do so in similar ways, and what are the
differences? A rise in quotations can be interpreted a posteriori as the result
of a speculative bubble but could be mere euphoria. How this does lead to
a rupture of the trend? Can the duration differences be interpreted? Can we
find universality classes?
Physics-like model of fracture or other phase transitions, including perco-
lation can be turned into some economic advantage. Along the same lines of
thought, the question was already touched upon in [4] within the sand pile
model. This allows not only a verbal analogy of index rupture in terms of
Patterns, Trends and Predictions in stock market 3
sand avalanches, but also some insight into the mechanisms. Through phys-
ical modeling and an understanding of parameters controlling the output,
as in the sand pile model, symptoms can be measured, whence to suggest
remedies is not impossible.
Another question raised below is the post-crash period. One might expect
from a physics point of view that if a crash looks like a phase transition, and
is characterized by scaling laws, as we will see it sometimes occurs [5,6], it
might be expected that a relation exists between amplitudes and laws on
both sides of the crash day [7]. As mentioned above, the crash might be
occurring on various days, with different breaks. It might be possible that
between drops some positive surge might be found. Thus some sorting of
behaviors into classes should be made as well. In fine, some discussion on
the foreign exchange market will be given in order to recall the detrended
fluctuation analysis method, so often used nowadays. It is applied below to
the U SD exchange rate vs. the (ten) currencies forming the EU R on Jan. 01,
2000, over a time interval including the most recent years. The observation
of the time variation of the power law scaling exponent of the DFA function
is shown to be correlated to crash time occurrence.
1.1 Tulipomania
In 1559, the first tulip bulb (TB) was brought to Holland from China by
Conrad Guenster [8]. In 1611, the tulip bulbs (TBs) were stocked and sold
on markets. In 1625, one tulip bulb was worth 5 dutch gulden (NLG). The
flower was considered so rare that wealthy aristocrats and merchants tripped
over themselves to buy the onions. Speculation ensued and the TBs became
wildly overvalued. The TBs were not necessarily planted, but were just stored
in the house salon. In 1635, 1 TB was worth 4 tons of wheat + 4 oxen + 8
tons rye + 8 pigs + one bed + 12 sheep + clothes + 2 wine casks + 4 tons
beer + 2 tons butter + 1000 pounds cheese + 1 silver drinking cup. In 1637, 1
TB was worth 550 NLG. One average house was worth 17 800 NLG, whence
about 30 TBs. However within 1637, over a 6 week time span the price of 1
TB went down 90 %.
In view of the shock, remedies had to be found and people called upon the
Amsterdam Parliament for legislation. It was decided that all contracts would
be void if they were dealt before Nov. 1636, and after that date the contracts
kept a 10 % value. Under some protest, people appealed to the Netherland
Supreme Court which ruled that this business of selling/buying TBs was mere
gambling, and no debt could be defined ”by law” nor ruled upon. Nowadays
one TB is worth 0.5 EUR. Too bad for long term investment strategies. The
TB became the classical example for illustrating the Extraordinary Popular
Delusions and the Madness of Crowds as described by Charles MacKay [9].
Just for the sake of physics history, let it be recalled that 1639 was the year
in which Galileo Galilei (Pisa, Feb. 15, 1564; Arcetri, Jan. 8, 1642) betrayed
science in saving his life.
4 M. Ausloos and K. Ivanova
1.2 Monopolymania
Another set of financial crises is that of the Compagnie du Mississipi [10] and
that of the South Sea Company [11,12]. In 1715, John Law, a scot gambler,
had persuaded Philippe, Regent of France, to consider a banking scheme
that promised to improve the financial condition of the kingdom. In theory a
private affair, the system was linked from the beginning with liquidating the
national debt. When the monopoly of the Louisiana trade was surrendered in
1717, Law created a trading company known as the Compagnie d’Occident (or
Compagnie du Mississipi) linked to the Royal Bank of France (first chartered
in 1716 as Banque Générale) and in which government bills were accepted
for the purchase of shares.
Law gained a monopoly on all French overseas trade. The result was a
huge wave of speculation as the value of a share went from its initial value,
i.e. 500 livres to 18 000 livres. When the paper money was presented at the
bank in exchange for gold, which was unavailable, panic ensued, and shares
felt by a factor of 2 in a matter of days.
In England, the Whigs represented the mercantile interests which had
profited from the War of the Spanish Succession War (1703-1711), and made
large profits by financing it, in doing so had created a National Debt which
had to be financed by further taxation. During the wars the government han-
dled more money than ever before in history, and they skimmed off a lot
through various methods, including the invention of the Bank of England in
1694. The South Sea Company was formed in 1711 by the Tory government
of Harley to trade with Spanish America, and to offset the financial support
which the Bank of England had provided for previous Whig governments.
They had in mind to establish a system like the Compagnie du Mississipi
Monopoly, [10] using the same sort of trading privileges and monopolies, those
granted to Britain after the Treaty of Utrecht. King George I of Great Britain
became governor of the company in 1718, creating confidence in the enter-
prise, which was soon paying 100 percent interest. In 1720 a bill was passed
enabling persons to whom the government owed portions of the national debt
to exchange their claims for shares in company stocks, and to become stock
holders. In the 1719-20 the England Public Debt went to South Sea Com-
pany stock holders, as approved by Parliament. On March 1 the stocks were
valued GBP 175, moved quickly to 200. Shortly the directors of the South
Sea Company had assumed three-fifths of the national debt. The company
expected to recoup itself from expanding trade, but chiefly from the foreseen
rise in the value of its shares. On June 1 the shares were valued 500 and
more than 1,000 in August 1720. Those unable to buy South Sea Company
stocks were inveigled by overly optimistic company promoters or downright
swindlers into unwise investments. Speculators took advantage of investors
to obtain subscriptions for sensibly unrealistic projects. By September 1720
however, the market had collapsed, and by December 1720 South Sea Com-
pany shares were down to 124, dragging others, including government stocks
Patterns, Trends and Predictions in stock market 5
with them. Many investors were ruined, and the House of Commons ordered
an inquiry, which showed that ministers had accepted bribes and speculated.
From a physics point of view let it be recalled that I. Newton (Woolsthorpe,
Dec. 25, 1642; London, March 20, 1727) invested in such South Sea Company
stocks and lost quite a bit of money [13].
1.3 WallStreetmania
In the years from 1925 to 1929 one could easily go to a broker and purchase
stocks on margin, i.e. instead of buying stocks with real cash money, one could
purchase them with some cash down and the rest on credit. The Coolidge
administration had a laissez-faire policy, i.e. a government policy of non-
intervention. It was almost la façon de vivre to play in the stock market [14].
That allowed a speculation bubble to grow unchecked. The Federal Reserve
powers on economic matters were not utilized as could be done nowadays.
Many successions of mini crashes and rallies began as early as March 1929.
The summer of 1929 hearkened somewhat of the good old days of optimism.
The market appeared to be stable. On Sept. 3, a bear market became firmly
established, and on Thursday Oct. 24, 1929 the famous 1929 crash occurred
(Fig. 1).
Fig. 1. Famous Wall Street Crashes : 1929, 1987, 1997; taken from color web site
http://www.lowrisk.com/crash/87vs97.htm; from top to bottom on day 201 : blue
(97), red (29), black (87); on day 700 : red (29), black(87), blue(97)
The 1987, 1997, and more recent 1999, 2000, 2001 crashes are reminis-
cent and even copies of the above ones (Fig. 1). The symptoms look similar :
artificially built euphoria, malingnantly established speculation, easy access
6 M. Ausloos and K. Ivanova
150
clustering
100
50
DAX volatility
−50
−100
−150
1985 1986 1987 1988 1989
date
Fig. 2. DAX volatility between Jan. 01, 85 and Dec. 31, 88
8000
7000
6000
5000
t*
DAX
4000
t
c
3000 t*
2000
1000 t
c
0
1980 1984 1988 1992 1996
date
Fig. 3. DAX evolution between Jan.01, 1980 and Dec.31, 1998 with the mean field
behavior, the time(s) t∗ , corresponding to the Ginzburg − Levanyuk temperature
bounding the critical fluctuation region, and the critical crash day(s) tc
8 M. Ausloos and K. Ivanova
The law for y(t) diverges at t = tc if m > 0. This evolution is decorated with
oscillations converging at the rupture point tc . This law is similar to that of
critical points, and generalizes the situation for cases in which a hierarchical
lattice structure exists, in other words a Discrete Scale Invariance (DSI) is
subjacent [29].
The relationship (1) has been proposed elsewhere in order to fit experi-
mental measurements of sound wave rate emissions prior to the rupture of
heterogeneous composite stressed up to failure [30]. The same type of com-
plex power law behavior has been also observed as a precursor of the Kobe
earthquake in Japan [31]. Such log-periodic corrections have been recently
reported in biased diffusion on random lattices [32].
As early as April 1997, Vandewalle and Ausloos performed a series of
investigations in order to emphasize crash precursors [33,34,35]. The closing
values of the Dow Jones Industrial Average (DJIA) and the Standard & Poor
500 (S&P500) were used for tests. A law slightly different from Eq.(1) was
proposed [33,36]. A strong indication of a so-called crash event or market
rupture point was numerically discovered [33,34,35,36]. Further data analysis
(in Aug. 97) [37] including a risk measure [38] indicated a crash to occur
in between the end of October 1997 and mid-November 1997. The crash
occurred effectively on Monday October 27th, 1997 [36] !
Eventhough the crash of October 1997 was predicted [36,37], the scientific
(physics or economy) [39,40] and media [35] community is actually divided
between those who believe in such a crash prediction and those who believe
that crashes are unpredictable events and such findings were mere luck or
at best accidental [27,28,41]. We discuss a little bit more the predictability
problem and findings in this paper going beyond a previous report [4].
Patterns, Trends and Predictions in stock market 9
1
Oct 16, 89
Oct 19, 87
Aug 19, 91
Oct 26, 87
May 29, 62
Oct 28, 87
0.5
Oct 28, 97
y(i)/y(0) −1
−0.5
12 12
10 10
(a) Oct. 16, 89 (b) Oct. 19, 87
10 10
10 10
8 8
10 10
S(f)
S(f)
6 6
10 10
4 4
10 10
2 2
10 10
0 0
10 −3 −2 −1 0
10 −3 −2 −1 0
10 10 10 10 10 10 10 10
−1
f (days ) f (days−1)
12 12
10 10
(c) Aug. 19, 91 (d) Oct. 26, 87
10 10
10 10
8 8
10 10
S(f)
S(f)
6 6
10 10
4 4
10 10
2 2
10 10
0 0
10 −3 −2 −1 0
10 −3 −2 −1 0
10 10 10 10 10 10 10 10
−1 −1
f (days ) f (days )
12 12
10 10
(e) May 29, 62 (f) Oct. 28, 87
10 10
10 10
8 8
10 10
S(f)
S(f)
6 6
10 10
4 4
10 10
2 2
10 10
0 0
10 −3 −2 −1 0
10 −3 −2 −1 0
10 10 10 10 10 10 10 10
f (days−1) f (days−1)
Fig. 5. The power spectrum of the 600 day DAX index evolution signal correspond-
ing to the six major crashes between Oct. 01, 59 and Dec. 31, 96
The roughness behavior [52,53,54] of the DAX index evolution signal be-
fore crashes can be defined trough the fractal dimension D of the signal, i.e.
[52]
3−β
D=E+ , (7)
2
where E is the Euclidian dimension. The values of β and D are reported in
Table 1 with the crash dates and relative amplitude of the 6 major DAX
crashes which occurred between Oct. 01, 1959 and Dec. 30, 1996. The same
12 M. Ausloos and K. Ivanova
type of data is reported in Table 2 for the 3 major DAX crashes in October
1997. The power spectrum of the large Oct. 28, 97 crash is shown in Fig. 6.
12
10
Oct. 28, 97
10
10
8
10
S(f)
6
10
4
10
2
10
0
10 −3 −2 −1 0
10 10 10 10
−1
f (days )
Fig. 6. The power spectrum of the 600 day DAX index evolution signal prior to
the Oct. 28, 97 crash
Table 1. Crash dates and relative amplitude of the 6 major DAX crashes having
occurred between Oct. 01, 1959 and Dec. 30, 1998; power law exponent β− of the
signal power spectrum (600 data points) and corresponding fractal dimension D−
of the signal prior to, and β+ and D+ , after these 6 major DAX crashes.
Table 2. The 3 major DAX crashes in October 1997 : crash dates, relative ampli-
tude, power law exponent β− of the signal power spectrum (600 data points prior
to these 3 major DAX crash dates) and the corresponding fractal dimension D− .
1
Oct 16, 89
Oct 19, 87
Aug 19, 91
Oct 26, 87
May 29, 62
Oct 28, 87
0.5
Oct 28, 97
y(i)/y(0) −1
−0.5
0 1 2 3
10 10 10 10
Time after the crash (days)
Fig. 7. The DAX recovery signal evolution after the six strongest DAX crashes
having occured between Jan. 01, 85 and Dec. 31, 88, and after the strongest DAX
crash in 1997 ; y0 denotes the index value at the closing of the crash day
The DFA method [60] consists in dividing the whole data sequence y(n) of
length N into N/τ non overlapping boxes, each containing τ points. Then,
the local trend
z(n) = an + b (8)
Patterns, Trends and Predictions in stock market 15
12 12
10 10
(a) Oct. 16, 89 (b) Oct. 19, 87
10 10
10 10
8 8
10 10
S(f)
S(f)
6 6
10 10
4 4
10 10
2 2
10 10
0 0
10 −3 −2 −1 0
10 −3 −2 −1 0
10 10 10 10 10 10 10 10
−1
f (days ) f (days−1)
12 12
10 10
(c) Aug. 19, 91 (d) Oct. 26, 87
10 10
10 10
8 8
10 10
S(f)
S(f)
6 6
10 10
4 4
10 10
2 2
10 10
0 0
10 −3 −2 −1 0
10 −3 −2 −1 0
10 10 10 10 10 10 10 10
−1 −1
f (days ) f (days )
12 12
10 10
(e) May 29, 62 (f) Oct. 28, 87
10 10
10 10
8 8
10 10
S(f)
S(f)
6 6
10 10
4 4
10 10
2 2
10 10
0 0
10 −3 −2 −1 0
10 −3 −2 −1 0
10 10 10 10 10 10 10 10
f (days−1) f (days−1)
Fig. 8. The power spectrum of the 600 day DAX index evolution signal after the
six major crashes having occured between Oct. 01, 59 and Dec. 31, 96
Averaging F (τ ) over the N/τ intervals gives a function depending on the box
size τ . The above calculation is repeated for different box sizes τ . If the y(n)
data are randomly uncorrelated variables or short range correlated variables,
the behavior is expected to be a power law
hF (τ )2i1/2 ∼ τ α (10)
with an exponent 1/2 [60] if the excursion is governed by a mere random walk.
An exponent α 6= 1/2 in a certain range of τ values implies the existence of
LRPLC in that time interval. Mathematically, the correlation of a future
increment y(n) − y(0) with a past increment y(0) − y(−n) is given by
h(y(0) − y(−n))(y(n) − y(0))i
Γ (n) = = 22α−1 − 1, (11)
h(y(n) − y(0))2 i
where the correlations are normalized by the variance of y(n). For α > 1/2,
there is persistence, i.e. Γ > 0. In this case, if in the immediate past the signal
has a positive increment, then on the average an increase of the signal in the
immediate future is expected. In other words, persistent stochastic processes
exhibit rather clear trends with relatively little noise. An exponent α < 1/2
means antipersistence, i.e. Γ < 0. In this case, an increasing value in the
immediate past implies a decreasing signal in the immediate future, while
a decreasing signal in the immediate past makes an increasing signal in the
future probable. In so doing, data records with α < 1/2 appear very noisy
(rough). The α = 0 situation corresponds to the so-called white noise. Finally,
one should note that α is nothing else than Ha, the so-called Hausdorff
exponent for fractional Brownian motions [53,54]. It can be useful to recall
[54] that the power spectrum of such random signals is characterized by a
power law with an exponent β = 2α − 1.
ATS
11 BEF
DEM
10 ESP
FIM
9 FRF
IEP
8 ITL
NLG
7 PTE
6 EUR
2
/ USD
1
1993 1994 1995 1996 1997 1998 1999 2000 2001 2002
hF (t)2i1/2 is shown for the whole data of Fig. 9. Moreover we plot the result
for a false EU R, i.e. a linear combination of the ten currencies forming the
EU R [64,65,66,67] Except for IEP , the functions are very close to a power
law with an exponent α = 0.51 ± 0.02 holding over two decades in time,
i.e. from about one week to two years. This finding clearly shows the non-
existence of LRPLC in the foreign exchange market with respect to the U SD.
Other cases showing marked deviations from Brownian motion have been
discussed elsewhere [65,67,68,69]. It can then be observed that a wide variety
of behaviors is found in the foreign currency exchange market. Exponent
values and the range over which a power law holds drastically vary from a
currency exchange rate to another. It appears that the currency exchange
rates can be classified into three different categories from the LRPLC point
of view.
First, the rates which exhibit an exponent α larger than 1/2 (persistent
behavior). This case corresponds to currency exchange rates between leading
currencies (e.g., U SD, JP Y , EU R) and so called weaker ones [69,70].
18 M. Ausloos and K. Ivanova
0
10
−1
10
−2
10 ATS
BEF
<F2(τ)>1/2
−3 DEM
10
ESP
−4
10 FIM
FRF
−5
10 IEP
ITL
−6 NLG
10 / USD
PTE
−7
EUR
10 0 1 2 3
10 10 10 10
τ
Fig. 10. Log-log plot of the DFA function showing how to obtain the α exponent
for the 11 exchange rates of interest for EU R/U SD. The fit slope being only of
interest, the DFA function data has been arbitrarily displaced along the vertical
axis
the data by some step toward the right along the financial sequence and α is
again calculated. Iterating this procedure one obtains a ”local measurement”
of the degree of ”long-range correlations” over T . It is crucial to choose the
most adequate box size T , i.e. to choose T of the same order of magnitude
as the maximum range τ over which the above power law is valid.
12
11 ATS
10 BEF
9 DEM
8 ESP
7 FIM
FRF
α
5 IEP
4 ITL
3 NLG
2 PTE
1 EUR
/ USD
1994 1995 1996 1997 1998 1999 2000 2001
τ
Fig. 11. Time dependence of the DFA local α-exponent for EU R and each currency
forming the EU R exchange rate with respect to U SD. The α-values are artificially
multiplied by two and then displaced along the vertical axis in order to make the
fluctuations noticeable. For each time dependent α a horizontal dashed line is drawn
to indicate a reference to Brownian fluctuations
analysis of the Dow Jones Industrial stock index around the 1987 October
crash was performed [71]. A similar pattern is found.
Other XR time series have been examined in order to check the non-
stationarity of α [65,67]. This does support the idea that the foreign currency
exchange markets are mainly governed by random conditions [72] or is said
to be efficient in more usual economic language. However, this unconditional
randomness cannot be extrapolated to speculating times nor emerging cur-
rencies [73]. Different universality classes thereby emerge. It may be useful to
recall that Hartmann[74] has examined the competition between U SD and
EU R in a more general (political and economy) framework.
4 Conclusions
The DAX has been analyzed from the point of view of crashes, in particular
the correlations in the signal volatility, before and after the critical days.
The search for the crash day is separated into two numerical problems, that
of the index divergence itself and that of the index oscillation frequency
acceleration on the other hand. By considering the envelope of the DAX, we
have demonstrated that before crashes, a log-periodic pattern exists. Even
though error bars are intrinsically large, it is surprising to see that a rupture
point is easily predicted. A hierarchical structure close to a fractal percolation
backbone (or tree) seems intrinsic at crashes. The stability of this result
should be tested in real time for the best future of our economic system.
A few foreign exchange currency rates with respect to the U SD have been
examined in order to illustrate the DFA technique, the intrinsic structure of
the DFA exponent, and its implications with respect to crashes.
Acknowledgements
Luc T. Wille is gladly thanked for inviting us to present the above results
and considerations, and enticing us into writing this report. Thanks to the
State of Florida for some financial support allowing the authors to participate
in the conference. MA thanks N. Vandewalle for numerous discussions.
References
1. B. Lauterbach, U. Ben-Zion: J Finance 48, 1909 (1993)
2. R. Roll: ’The International Crash of October 1987’, R. W. Kamphuis et al.,
eds.: Black Monday and the Future of Financial Markets, (Dow Jones/Irwin,
Homewood, IL, 1989)
3. N. Vandewalle, Ph. Boveroux, F. Brisbois: Eur. J. Phys. B 15, 547 (2000)
Patterns, Trends and Predictions in stock market 21
73. K. Ivanova, M. Ausloos: Eur. Phys. J. B 8, 665 (1999); Err. 12, 613 (1999)
74. Ph. Hartmann: Currency Competition and Foreign Exchange Markets. The Dol-
lar, the Yen and the Euro (Cambridge Univ. Press, Cambridge, 1998)
75. http://lowrisk.com/crash/crashcharts.htm; http://lowrisk.com/crash/87vs97.htm;
http://lowrisk.com/crash/1929crash.htm